monitor-training-with-tmux

Pass

Audited by Gen Agent Trust Hub on Mar 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses subprocess.run in tmux_wait.py and SKILL.md to execute tmux list-panes, checking the status of active processes within a session.- [COMMAND_EXECUTION]: The skill instructions in SKILL.md involve executing system diagnostic commands such as nvidia-smi, free -h, df -h, and dmesg | tail -50 to gather environment context when anomalies are detected.- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it ingests and analyzes data from external sources (tmux session logs).
  • Ingestion points: The skill captures up to 2000 lines of tmux session history using tmux capture-pane -t <session> -p -S -2000.
  • Boundary markers: No boundary markers or delimiters are specified to separate the captured logs from the agent's instructions.
  • Capability inventory: The skill has the capability to execute system commands and suggest fixes based on log analysis.
  • Sanitization: There is no evidence of sanitization or filtering of the captured logs before they are processed by the agent.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 13, 2026, 09:08 AM